Intelligent decision-making method based on new energy ship multi-dimensional risk coupling modeling and related equipment

Abstract

The invention provides an intelligent decision-making method based on multi-dimensional risk coupling modeling of new energy ships and related equipment. The method comprises the following steps: fusing multi-modal data of each new energy ship to obtain a target multi-modal feature vector; mining risk implicit association strength by using a large language model, and constructing a risk knowledge graph; performing risk time sequence evolution prediction by adopting a dynamic Bayesian network based on the atlas to obtain a single-ship risk prediction result; constructing a graph structure according to the single ship risk and the operation parameters, and carrying out space coupling modeling by adopting a graph convolutional network to obtain a cluster risk prediction result; and generating an optimal operation and maintenance strategy by adopting reinforcement learning based on a cluster risk prediction result and a reward function by taking a high-fidelity digital twinborn body as a virtual environment. Therefore, according to the method, the problem of closed-loop adaptive control from risk deduction to decision response in a complex navigation scene is effectively solved by constructing a unified modeling mechanism of multi-dimensional risk coupling and linking real-time control strategy generation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of new energy ship safety operation, and particularly relates to an intelligent decision-making method based on multi-dimensional risk coupling modeling of new energy ships and related equipment. BACKGROUND

[0002] With the transformation of the shipbuilding industry towards green and intelligent, new energy ships such as pure battery and fuel cell have become the mainstream of the shipping industry, but their safety operation faces the complex challenges of three-dimensional risks of "environment-power-operation". The navigation scenarios cover near sea, inland river and coastal areas, and the meteorological and hydrological conditions and channel conditions are significantly different in different scenarios. The risk factors show strong coupling, time dynamics and scenario dependence characteristics, and a single risk factor can easily cause systemic safety accidents through chain conduction. Therefore, building a full-dimensional risk dynamic coupling model to realize accurate characterization, real-time prediction and intelligent decision-making of risks is the core requirement of new energy ship safety operation.

[0003] Existing new energy ship risk modeling technologies mostly use static analysis methods, which are difficult to fully represent the complex causal relationship between environmental, power and operation risk factors and their chain conduction mechanism. The risk state continuously evolves with the navigation process, and the time variation characteristics of environmental parameters such as swell and water level are interwoven with the gradual trend of power system performance. Static evaluation at a single moment cannot reflect the cumulative changes of risks and cannot accurately predict short-term sudden risks. At the same time, the development of risk response strategies mainly relies on artificial experience judgment, and the response link between early warning information and operation adjustment is long, which is difficult to adapt to the rapid changes of risks in the navigation process.

[0004] Therefore, there is still a lack of unified modeling mechanism for the dynamic coupling evolution of multi-dimensional risk factors, and there is no adaptive generation capability of real-time control strategies linked thereto, making it difficult to realize the closed-loop adaptive control from risk deduction to decision response in complex navigation scenarios. SUMMARY

[0005] The purpose of the present application is to at least solve one of the above technical defects, in particular the lack of a unified modeling mechanism for the dynamic coupling evolution of multi-dimensional risk factors, and the lack of adaptive generation capability of real-time control strategies linked thereto, making it difficult to realize the closed-loop adaptive control from risk deduction to decision response in complex navigation scenarios.

[0006] In a first aspect, the present application provides an intelligent decision-making method based on multi-dimensional risk coupling modeling of new energy ships, the method comprising:

[0007] In a new energy ship fleet, the multi-modal data of each new energy ship is fused to obtain a target multi-modal feature vector of each new energy ship;

[0008] For each new energy vessel, based on its target multimodal feature vector, a large language model is used to mine the strength of implicit risk associations among multidimensional risk factors, and a risk knowledge graph is constructed according to the strength of implicit risk associations.

[0009] Based on the risk knowledge graph of each new energy vessel, a dynamic Bayesian network is used to predict the temporal evolution of risk, and the individual risk prediction results of each new energy vessel are obtained.

[0010] Using the individual risk prediction results and operating status parameters of each new energy vessel as node features, a graph structure is constructed. Spatial coupling modeling is performed using a graph convolutional network to obtain the cluster risk prediction results for each new energy vessel.

[0011] Using a high-fidelity digital twin of a new energy vessel as a virtual simulation environment, and based on the risk prediction results of each cluster and the preset reward function, a reinforcement learning algorithm is used to generate the optimal vessel operation and maintenance strategy.

[0012] In one embodiment, the step of mining the strength of implicit risk associations among multidimensional risk factors using a large language model based on its target multimodal feature vector includes:

[0013] The strength of the implicit association of risk is determined by the following expression:

[0014]

[0015] in, , Let i and j be the i-th and j-th risk factor nodes in the multidimensional risk factors. The strength of the association between risk factors For large language models, For the target multimodal feature vector, As a contextualized prompt template, It is a normalized exponential function.

[0016] In one embodiment, the steps of using a dynamic Bayesian network to predict the temporal evolution of risk based on the risk knowledge graph of each new energy vessel, and obtaining the individual vessel risk prediction result for each new energy vessel, include:

[0017] The individual risk prediction result for each new energy vessel is determined according to the following expression:

[0018]

[0019] in, This represents the risk state vector of multidimensional risk factors at time t in the risk knowledge graph. The temporal feature vector representing the target multimodal feature vector. This represents the posterior probability of the risk state at time (t+1) due to the multidimensional risk factors, which is also the single-ship risk prediction result. Represents the state transition matrix. This represents a multilayer perceptron. This represents the Sigmoid function.

[0020] In one embodiment, the steps of constructing a graph structure using the individual risk prediction results and operational status parameters of each new energy vessel as node features, and employing a graph convolutional network for spatial coupling modeling to obtain the cluster risk prediction results for each new energy vessel include:

[0021] The cluster risk prediction result for each new energy vessel is determined according to the following expression:

[0022]

[0023] in, This represents the graph structure feature vector, which is also the cluster risk prediction result for each new energy vessel. Represents a graph convolutional network. This represents an adjacency matrix. The element values ​​of the adjacency matrix are assigned corresponding base values ​​according to the different intervals in which the distances between the new energy vessels fall, and the base values ​​are weighted and corrected according to the speed differences of the new energy vessels. This represents the node feature matrix, which contains the individual risk prediction results and operational status parameters for each new energy vessel. This represents the Sigmoid function. Represents the normalized adjacency matrix. This represents the activation function. and These represent the trainable weight matrices, respectively.

[0024] In one embodiment, the method further includes:

[0025] The risk knowledge graph is dynamically updated using the following expression:

[0026]

[0027] in, This represents the state of the risk knowledge graph at time t. Indicates the update step size. This indicates the strength of the implicit risk association among the multidimensional risk factors generated by the large language model. This indicates the confirmation result of the content generated by the large language model. This represents the cosine similarity.

[0028] In one embodiment, the method further includes:

[0029] By using federated learning and collaborating local data from multiple new energy ship data sources, the parameters of dynamic Bayesian networks and graph convolutional networks are trained, and the general model parameters of dynamic Bayesian networks and graph convolutional networks are obtained.

[0030] Scene data of the target navigation scenario is acquired, and the general model parameters of the dynamic Bayesian network and the graph convolutional network are fine-tuned through transfer learning to obtain the adapted model parameters of the dynamic Bayesian network and the graph convolutional network in the target navigation scenario.

[0031] In one embodiment, the method further includes:

[0032] The SHAP algorithm was used to determine the importance values ​​of various risk factors in the multidimensional risk factors.

[0033] The importance values ​​of various risk factors and the risk knowledge graph are input into a domain-fine-tuned large language model to generate natural language explanations and operational suggestions that fit the operation and maintenance scenario.

[0034] Secondly, this application provides an intelligent decision-making device based on multi-dimensional risk coupling modeling of new energy ships, the device comprising:

[0035] The multimodal data fusion module is used to fuse the multimodal data of each new energy vessel in the new energy vessel fleet to obtain the target multimodal feature vector of each new energy vessel.

[0036] The risk knowledge graph construction module is used to mine the strength of implicit risk associations between multidimensional risk factors for each new energy vessel based on its target multimodal feature vector and a large language model, and to construct a risk knowledge graph based on the strength of the implicit risk associations.

[0037] The single-ship risk prediction result determination module is used to predict the risk time series evolution based on the risk knowledge graph of each new energy vessel and to obtain the single-ship risk prediction result for each new energy vessel.

[0038] The cluster risk prediction result determination module is used to construct a graph structure using the individual ship risk prediction results and operating status parameters of each new energy vessel as node features, and to perform spatial coupling modeling using a graph convolutional network to obtain the cluster risk prediction result for each new energy vessel.

[0039] The optimal ship operation and maintenance strategy generation module is used to generate the optimal ship operation and maintenance strategy based on the risk prediction results of each cluster and the preset reward function, using a high-fidelity digital twin of a new energy ship as a virtual simulation environment and a reinforcement learning algorithm.

[0040] Thirdly, this application provides a storage medium storing computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the steps of any of the intelligent decision-making methods based on multidimensional risk coupling modeling of new energy ships as described in the above embodiments.

[0041] Fourthly, this application provides a computer device, including: one or more processors, and a memory;

[0042] The memory stores computer-readable instructions, which, when executed by one or more processors, perform the steps of any of the intelligent decision-making methods based on multidimensional risk coupling modeling of new energy ships as described in the above embodiments.

[0043] As can be seen from the above technical solutions, the embodiments of this application have the following advantages:

[0044] This application provides an intelligent decision-making method and related equipment based on multidimensional risk coupling modeling for new energy vessels. First, by fusing multimodal data for each new energy vessel, a target multimodal feature vector is obtained, achieving a comprehensive and unified representation of the vessel's operational status and providing a precise data foundation for multidimensional risk coupling analysis. Then, for each new energy vessel, based on its target multimodal feature vector, a large language model is used to mine the implicit risk correlation strength between multidimensional risk factors, and a risk knowledge graph is constructed based on the implicit risk correlation strength, thereby revealing the deep coupling mechanism between risk factors and establishing a semantic framework for unified multidimensional risk modeling. On this basis, based on the risk knowledge graph of each new energy vessel, a dynamic Bayesian network is used to predict the temporal evolution of risks, obtaining the single-level risk knowledge graph of each new energy vessel. The ship risk prediction results enable the quantitative deduction of the dynamic evolution law of risk. Furthermore, using the individual ship risk prediction results and operational status parameters of each new energy vessel as node features to construct a graph structure, and employing a graph convolutional network for spatial coupling modeling, the cluster risk prediction results for each new energy vessel are obtained. This extends individual ship risk to the fleet cluster level, characterizing the risk propagation and coupling effects under the mutual influence of ships, forming a unified model for the spatiotemporal evolution of multidimensional risks in complex navigation scenarios. Finally, using a high-fidelity digital twin of the new energy vessel as a virtual simulation environment, based on the risk prediction results of each cluster and a preset reward function, a reinforcement learning algorithm is used to generate the optimal ship operation and maintenance strategy. This achieves adaptive linkage between risk deduction results and real-time control strategies, completing closed-loop adaptive control from risk perception and evolution prediction to strategy generation. Therefore, this method, by constructing a unified modeling mechanism for multidimensional risk coupling and linking it with real-time control strategy generation, effectively solves the closed-loop adaptive control problem from risk deduction to decision response in complex navigation scenarios, significantly improving the safety and intelligence level of new energy vessel operation. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 A flowchart illustrating the intelligent decision-making method based on multidimensional risk coupling modeling of new energy ships provided in this application embodiment;

[0047] Figure 2 A schematic diagram of the structure of an intelligent decision-making device based on multidimensional risk coupling modeling of new energy ships provided in an embodiment of this application;

[0048] Figure 3 This is a schematic diagram of the internal structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0049] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0050] This application provides an intelligent decision-making method based on multi-dimensional risk coupling modeling of new energy ships. The following embodiments illustrate this method using a computer device as an example. It is understood that the computer device can be any device with data processing capabilities, including but not limited to a single server, server cluster, personal laptop, desktop computer, etc. Figure 1 As shown, the method may include the following steps:

[0051] S101: In the new energy vessel fleet, the multimodal data of each new energy vessel are fused to obtain the target multimodal feature vector of each new energy vessel.

[0052] In this application, a new energy vessel fleet refers to a navigation formation or operational cluster consisting of two or more new energy vessels. New energy vessels refer to vessels that use clean energy sources such as electricity, hydrogen energy, and liquefied natural gas as their power source. Multimodal data refers to heterogeneous data acquired from various types of sensors and data acquisition systems deployed on new energy vessels, including but not limited to vessel navigation status data, power system operation data, environmental perception data, and equipment health status data. Fusion refers to the process of aligning, associating, and comprehensively processing the above-mentioned multimodal data to form a unified representation. The target multimodal feature vector refers to a high-dimensional feature vector generated after fusion processing that can comprehensively characterize the current comprehensive state of the new energy vessel.

[0053] In the specific implementation process, for each new energy vessel in the fleet, a standardized specification for multimodal data of new energy vessels is first formulated, clarifying the storage format and preprocessing requirements of various types of data, laying the foundation for cross-source data fusion. Specifically, power system sensor data, including parameters such as battery voltage, motor temperature, and liquefied natural gas pressure, is stored in a time-series database format to accurately capture the dynamic trends of parameter changes. Automatic Identification System (AIS) trajectory data, including latitude, longitude, speed, heading, and channel position information, is stored in a spatial coordinate format to adapt to subsequent spatial relationship modeling needs. Meteorological and hydrological data, covering information such as coastal swell height, inland river water level fluctuations, nearshore wind force, and visibility, is stored in a tagged format to simplify the feature matching process under different navigation scenarios. Text data such as accident cases and operation logs are stored in a structured text format, extracting key fault information and operation records to achieve standardized expression of unstructured knowledge.

[0054] Building upon this foundation, a data encoding model is constructed based on the multimodal Transformer. Its core function is to address the incommensurability problem of heterogeneous data. Through a unified mapping, it transforms data of different types and dimensions into homogeneous feature vectors while fully preserving the core information of each modality. In this model, sensor time-series features are extracted using Time2Vec encoding technology, forming a 768-dimensional feature vector. This encoding method effectively captures the dynamic dependencies between the time dimension and parameter states. The AIS trajectory spatial features are encoded using a Graph Convolutional Network (GCN), which also generates a feature vector with a dimension of 768. This is used to quantify spatial interactions during ship navigation. Text features are extracted using a BERTbase pre-trained model, transforming unstructured text knowledge into a computable vector of dimension 768. Meteorological and hydrological features are encoded using a visual language pre-trained model (VLM) to obtain an environmental feature vector with a dimension of 768. This achieves the fusion of environmental factors with other modal data from the same source. The above four types of feature vectors are combined into the final multimodal feature vector through a weighted fusion method. Its dimension is set to 1024, and the fusion formula is as follows:

[0055]

[0056] In the formula, the feature weights , , , The contribution values ​​for sensor features, spatial features, text features, and environmental features are respectively assigned. Initial values ​​are set to 0.35, 0.25, 0.2, and 0.2 based on domain experience, and are continuously optimized during model training through an adaptive adjustment mechanism to adapt to the feature fusion requirements of different navigation scenarios. The multimodal Transformer, as the core encoding framework, can deeply explore the implicit correlations between cross-modal data, such as the potential relationship between battery temperature changes and surge intensity, thereby avoiding risky misjudgments caused by the loss of single-modal information.

[0057] Simultaneously, a data cleaning and alignment model was constructed to ensure data quality. This model employs the 3σ criterion to remove outliers from sensor data, uses the K-nearest neighbor algorithm to fill in missing data, leverages semantic matching technology to achieve accurate alignment between text data and fault types, and uses temporal calibration to ensure the synchronization of timestamps between sensor data and AIS trajectory data. Through these processes, data noise and bias are effectively eliminated, ensuring accurate correlation of multi-source data, ultimately forming a standardized dataset Dtotal, providing high-quality data support for subsequent risk modeling and intelligent decision-making.

[0058] Multimodal data fusion is performed on each new energy vessel to obtain its own target multimodal feature vector. The aim is to systematically integrate heterogeneous information scattered across different types of sensors and data sources, thereby generating a unified representation that comprehensively reflects the overall state of a single vessel. This approach eliminates the semantic gap caused by differences in acquisition methods, physical meanings, and data structures between multimodal data, enabling subsequent analysis to be based on more complete and accurate information. Simultaneously, the fusion process uncovers deep-seated correlations that cannot be reflected in single-modal data, such as the potential correspondence between power system parameters and environmental fluctuations, thus providing a high-quality data foundation for the identification and coupled modeling of multidimensional risk factors. Furthermore, the generated target multimodal feature vectors have fixed dimensions and a standardized format, facilitating direct use by subsequent algorithms such as large language models, dynamic Bayesian networks, and graph convolutional networks. This significantly improves the efficiency of data processing and the stability of model training throughout the intelligent decision-making process, ultimately laying a solid data foundation for accurately perceiving the operational status of each vessel and achieving differentiated risk management.

[0059] S102: For each new energy vessel, based on its target multimodal feature vector, use a large language model to mine the strength of implicit risk associations between multidimensional risk factors, and construct a risk knowledge graph based on the strength of implicit risk associations.

[0060] In this application, the large language model is a deep learning model pre-trained on massive text data, possessing natural language understanding and generation capabilities, and able to extract semantic information and potential associations from input data; multidimensional risk factors are various potential risk elements affecting the safe operation of new energy ships, including but not limited to power system anomalies, changes in the navigation environment, and human operational errors; the implicit association strength of risks refers to the degree of potential mutual influence between different risk factors that is difficult to discover directly through explicit rules, and is represented by quantitative values; the risk knowledge graph is a knowledge base organized in the form of a graph structure, where nodes represent risk factors, edges represent the association relationships between risk factors, and the edges are assigned association strength attributes to characterize the coupling network between multidimensional risks.

[0061] In practical implementation, an independent processing channel is first established for each new energy vessel. Once the target multimodal feature vector for each new energy vessel is generated, it is input to the large language model service interface. The large language model receives this feature vector as input, which integrates multidimensional information such as sensor time-series data, ship automatic identification system trajectory data, text logs, and meteorological and hydrological information. Internally, the model uses a multi-layer attention mechanism to parse the feature vector and extract the implicit risk semantic representation. To guide the model to focus on the correlation mining between risk factors, a set of prompt templates is pre-designed. The feature vector is concatenated with the prompt text and input into the model. The prompt text content is, for example, "Please analyze the possible risk factors in the following ship status characteristics and evaluate the mutual influence strength between each risk factor." Based on its pre-trained knowledge base and understanding of the input features, the large language model outputs a structured risk correlation description, which includes a list of identified risk factors and the correlation strength value between each pair of risk factors. The correlation strength is represented by a floating-point number between 0 and 1.

[0062] Next, the structured results output by the large language model are analyzed. For each new energy vessel, all identified risk factors are extracted from the output and used as nodes in the knowledge graph. For each pair of related risk factors, a directed edge is created from the source factor to the target factor, and the correlation strength value is stored as an attribute of the edge. Each risk factor node also includes a quantified risk value of that factor in the current vessel state, which can be directly mapped from the corresponding component in the target multimodal feature vector. An independent risk knowledge graph instance is created in the database for each new energy vessel, using the vessel identifier as the namespace to ensure that the risk knowledge graphs of different vessels are isolated from each other. Whenever a new target multimodal feature vector arrives, the above process is repeated to incrementally update the output of the large language model. That is, newly identified risk factor nodes are added to the graph, newly discovered related edges are updated between existing nodes, and edge attributes are corrected if the correlation strength between existing nodes changes. Through this dynamic update mechanism, the risk knowledge graph can reflect the multidimensional risk coupling state of the vessel in real time.

[0063] Based on the target multimodal feature vectors of each new energy vessel, a large language model is used to mine the strength of implicit risk associations and construct a risk knowledge graph. Leveraging the powerful semantic understanding capabilities of the large language model, it can automatically identify hard-to-find implicit risk associations from multimodal fusion features, thus overcoming the subjectivity and limitations of manual experience-based analysis. The constructed risk knowledge graph presents the coupling network between multidimensional risk factors in an intuitive graph structure. It not only clearly shows the risk propagation path and key risk nodes, but also provides structured prior knowledge for subsequent temporal evolution prediction using dynamic Bayesian networks. This allows risk prediction to be based on real coupling relationships, significantly improving the comprehensiveness and accuracy of risk identification.

[0064] S103: Based on the risk knowledge graph of each new energy vessel, a dynamic Bayesian network is used to predict the temporal evolution of risk, and the single-ship risk prediction results of each new energy vessel are obtained.

[0065] In this application, the Dynamic Bayesian Network is a probabilistic graphical model that extends the time dimension. It can model the uncertainty process of the system state evolution over time by introducing time slices and state transition probabilities, and is suitable for risk time series prediction scenarios. Risk time series evolution prediction refers to the dynamic process of inferring the probability of occurrence and risk level of each risk factor in multiple future time slices based on the risk coupling state presented by the risk knowledge graph at the current moment, using the Dynamic Bayesian Network. The single-ship risk prediction result refers to the structured output generated independently for each new energy ship, which includes the predicted probability of various risk factors at future moments and the comprehensive risk rating.

[0066] In practical implementation, a dynamic Bayesian network model based on a risk knowledge graph is first constructed for each new energy vessel. The risk knowledge graph at the current moment is extracted from a graph database as the initial network structure. This graph contains multiple risk factor nodes, directed edges between nodes, and association strength attributes. The association strength attributes are transformed into initial values ​​for the conditional probability table in the dynamic Bayesian network; a higher association strength indicates a higher probability of influence of the parent node on the child node. Simultaneously, the quantified risk value of each risk factor node at the current moment is obtained from the target multimodal feature vector, serving as observational evidence for the initial time slice.

[0067] Next, the time parameters of the dynamic Bayesian network are set, with the time slice interval set to 10 minutes and the prediction window length set to 6 time slices, thus extrapolating the risk evolution trend over a 60-minute period. The state transition probability of the same risk factor between different time slices is statistically learned from historical navigation data and used to describe the natural evolution law of the risk factor. After completing the construction of the time-unfolded network, the clustered tree propagation algorithm is used for inference calculation. The observation evidence at the current moment is input into the network, and the posterior probability distribution of each risk factor node in each subsequent time slice is obtained through forward propagation. The inference results are output in the form of probability values, and the overall risk score for each future moment is calculated based on the centrality weight of each risk factor, forming a time series of single-ship risk prediction results, which is stored in a time series database for subsequent steps. To support the expansion of the fleet size, the inference tasks of different ships can be executed in parallel on multiple computing nodes to ensure real-time requirements.

[0068] Based on the risk knowledge graph of each new energy vessel, a dynamic Bayesian network is used for risk temporal evolution prediction. The aim is to extend the static risk coupling structure to the time dimension, using a probabilistic graphical model to characterize the dynamic evolution of risk factors over time. This implementation method allows risk prediction to move beyond static assessment at the current moment, quantifying the probability of occurrence of each risk factor and the overall risk level across multiple future timeframes. The dynamic Bayesian network fully utilizes the implicit relationships already discovered in the risk knowledge graph as network topology priors, ensuring that the temporal reasoning process is strictly based on real risk propagation paths, thus significantly improving the scientific rigor and accuracy of the prediction results. The final single-ship risk prediction results serve as an independent risk evolution profile for each vessel, providing a temporally aligned input foundation for subsequent cluster-level spatial coupling modeling, supporting a progressive analysis from single-point risk deduction to fleet risk perception.

[0069] S104: Using the individual risk prediction results and operating status parameters of each new energy vessel as node features, a graph structure is constructed. Spatial coupling modeling is performed using a graph convolutional network to obtain the cluster risk prediction results for each new energy vessel.

[0070] In this application, operational status parameters refer to the dynamic operational data of the new energy vessel at the current moment, including but not limited to real-time monitoring indicators such as vessel position, speed, heading, power system load rate, and equipment health index; node features refer to the high-dimensional feature vector formed by splicing and combining the risk prediction results of a single vessel with the operational status parameters, used to characterize the comprehensive state of the vessel in the graph structure; graph structure refers to a networked data model constructed with each vessel in the new energy vessel fleet as a node and the spatial distance, waterway connection relationship, or communication topology between vessels as edges; graph convolutional network is a deep learning model specifically for processing graph structure data, which can update the feature representation of the target node by aggregating neighbor node information, thereby capturing the spatial dependency relationship between nodes; spatial coupling modeling refers to the process of using graph convolutional networks to model the risk mutual influence relationship between vessels in the fleet, aiming to quantify how the risk state of a vessel is affected by the risk state of surrounding vessels; cluster risk prediction results refer to the final risk prediction output of each new energy vessel after spatial coupling modeling, considering the risk mutual influence of surrounding vessels.

[0071] In practical implementation, the first step is to construct a graph-structured data for spatial coupling modeling. Individual ship risk prediction results for all new energy vessels at the current moment are read from a time-series database. These results contain the risk probability sequence for each vessel across multiple predicted time slices. Simultaneously, the current operational status parameters of each vessel are obtained from real-time data streams, including latitude and longitude coordinates, current speed, heading, and key indicators of the power system. The individual ship risk prediction results and operational status parameters for each vessel are concatenated along the feature dimension to form a fixed-dimensional node feature vector. This vector simultaneously contains the vessel's future risk evolution trend and current operational status. Next, the edge connections between vessels need to be determined. The Euclidean distance between each pair of vessels is calculated based on their current positions. Vessel pairs with a distance less than a preset threshold (e.g., 5 nautical miles) are considered spatially adjacent, and an undirected edge is established between them. Channel topology constraints are also considered. For vessels within the same channel and traveling in the same direction, edge connections are established even if the spatial distance is slightly greater, to reflect the mutual influence within restricted waters. Edge connections are also established between vessels that establish data links through inter-ship communication systems. This forms a graph structure with ships as nodes and various relationships as edges, where the number of nodes equals the total number of ships in the current fleet.

[0072] After constructing the graph structure, the graph structure and its node features are input into a graph convolutional network (GCNN) for spatial coupling modeling. The GCNN adopts a two-layer architecture: the first layer aggregates the risk features of neighboring ships, and the second layer fuses the target ship's own features with the aggregated neighbor features. In each layer, the node features are updated by weighting the current node's features with the features of its neighboring nodes. The weights are determined by edge attributes such as the inverse of spatial distance. Then, a learnable weight matrix is ​​used for linear transformation, followed by a non-linear activation function such as ReLU for the output. After two layers of graph convolution, the output features of each node incorporate the risk status information of surrounding ships. To obtain the final cluster risk prediction result, the node features output by the GCNN are mapped to the same dimensional space as the single-ship risk prediction result through a fully connected layer, outputting a spatially coupled risk probability sequence and a comprehensive risk rating. This output is the cluster risk prediction result for each new energy vessel, reflecting the vessel's true risk level at future moments, considering the mutual influence of other ships in the fleet.

[0073] This study constructs a graph structure using the individual risk prediction results and operational status parameters of each new energy vessel as node features, and employs a graph convolutional network for spatial coupling modeling. The aim is to re-examine the previously isolated individual vessel risk prediction results within the overall framework of the fleet cluster. By explicitly modeling the spatial proximity relationships and mutual influence mechanisms between vessels, the risk prediction can accurately reflect the risk propagation and coupling effects in multi-vessel coexistence scenarios. The graph convolutional network aggregates the node features of neighboring vessels, integrating the risk status and operational dynamics of surrounding vessels into the feature update process of the target vessel. This effectively captures the dynamic relationship of how a vessel's risk changes are affected by neighboring vessels. Simultaneously, the operational status parameters, as components of the node features, ensure that the spatial coupling modeling is not only based on the risk prediction results but also incorporates the vessel's current real-time dynamic information, making the fused features more sensitive to sudden changes in the situation. The final cluster risk prediction results for each new energy vessel are corrected outputs after fully considering the mutual influence of all vessels within the fleet, significantly improving the accuracy and reliability of risk prediction in complex navigation scenarios and providing precise decision-making basis for subsequent development of operation and maintenance strategies involving multi-vessel collaboration.

[0074] S105: Using a high-fidelity digital twin of a new energy ship as a virtual simulation environment, and based on the risk prediction results of each cluster and the preset reward function, a reinforcement learning algorithm is used to generate the optimal ship operation and maintenance strategy.

[0075] In this application, the high-fidelity digital twin of a new energy vessel refers to a precise mapping model constructed in virtual space based on the physical entity of the new energy vessel. This model integrates the vessel's geometry, power system characteristics, kinematics and dynamics, and environmental interaction mechanisms, enabling it to respond to external inputs in real time in a manner consistent with the physical vessel. The preset reward function is a quantitative evaluation standard set for the reinforcement learning process. This function defines the immediate benefit or penalty for each decision action based on the cluster risk prediction results. For example, a positive reward is obtained for risk reduction, and a negative penalty is obtained for risk increase, while taking into account objectives such as operation and maintenance costs, energy efficiency, and task completion. The optimal vessel operation and maintenance strategy refers to a set of decision rules generated through extensive trial and error training in the digital twin environment using reinforcement learning algorithms. These rules guide the actual vessel to take the optimal operation sequence under different risk scenarios, including specific actions such as speed adjustment, route optimization, power system parameter setting, and collision avoidance.

[0076] Specifically, a high-fidelity digital twin of a new energy vessel is constructed to map the physical vessel's status and environmental changes in real time. Combined with reinforcement learning, the optimal operation and maintenance strategy is trained in a virtual environment to form a closed loop of modeling, prediction, decision-making, and feedback. This solves the problems of traditional decision-making relying on manual labor and having strong lag, and realizes proactive and intelligent risk prevention and control.

[0077] In the high-fidelity digital twin construction phase, it is necessary to comprehensively build the ship's geometric model, physical model, behavioral model, and environmental model. The geometric model recreates the ship's hull structure, power component layout, and battery compartment location based on ship design drawings. The physical model simulates the power system's operation and the ship's stress processes based on fluid mechanics and electrochemical theories. The behavioral model incorporates historical data such as crew operating habits and route planning preferences. The environmental model integrates real-time meteorological and hydrological data and waterway dynamic data to achieve real-time mapping of the physical ship's state to changes in its surrounding environment. To ensure that the digital twin can reliably replace the physical ship as a reinforcement learning training environment, model accuracy constraints need to be set. The simulation error calculation formula is:

[0078]

[0079] In the formula, This is for simulation error; For the state parameters of the digital twin; The parameters are measured from the physical ship; a 5% accuracy constraint ensures that the twin can replace the physical ship as a reinforcement learning training environment, avoiding simulation distortion that could lead to strategy failure.

[0080] In the reinforcement learning decision optimization phase, a high-fidelity digital twin is constructed as the virtual simulation environment. The agent is trained using the Proximal Policy Optimization (PPO) algorithm, which features high stability and sample efficiency, enabling rapid iterative learning of the optimal mapping relationship from risk states to decision actions. During training, the agent's observed state space includes the cluster risk prediction results for each ship, its own operating parameters, and surrounding environmental information. Decision actions encompass operational tasks such as speed adjustment, course deflection, and power system settings. The reward function is designed as follows:

[0081]

[0082] Among them, weight Take 0.5, Take 0.3, A weight of 0.2 is set to prioritize navigation safety. The total reward value, For safety rewards, calculations are performed using inversely normalized risk values; the lower the risk level, the higher the reward. The reward system, based on navigation efficiency and quantified by speed and maintenance time, aims to prevent excessive control measures from impacting normal operational efficiency. To penalize high-cost actions such as frequent emergency braking, a gradient ascent algorithm is used to continuously adjust the agent's policy parameters to maximize cumulative rewards, ultimately generating an optimal operation and maintenance strategy that balances safety, efficiency, and economy. Combined with the real-time mapping capabilities of the digital twin, a real-time linkage mechanism is formed, from risk warning to policy output, action execution, and finally, status updates.

[0083] Using a high-fidelity digital twin of a new energy vessel as a virtual simulation environment, and based on risk prediction results from various clusters and a preset reward function, a reinforcement learning algorithm is employed to generate the optimal vessel operation and maintenance strategy. The aim is to place the decision-making process within a virtual environment that realistically maps the physical state of the vessel, allowing the reinforcement learning agent to learn the optimal decision through repeated trial and error under risk-free conditions. The risk prediction results from each cluster serve as input to the environmental state, ensuring that the decision-making process fully considers the mutual influence between vessels and future risk evolution trends. The preset reward function provides an optimization objective for reinforcement learning, guiding the agent to generate strategies that balance safety and other operational requirements. Through interaction with the digital twin, the reinforcement learning algorithm continuously optimizes decision-making actions, ultimately generating an optimal vessel operation and maintenance strategy adaptable to different risk scenarios. This implementation achieves a closed-loop linkage from risk prediction to decision output, enabling the vessel operation and maintenance strategy to be adaptively generated based on real-time risk information, thus improving the intelligence level and response speed of decision-making.

[0084] In the above embodiments, firstly, by fusing the multimodal data of each new energy vessel, a target multimodal feature vector for each new energy vessel is obtained, achieving a comprehensive and unified representation of the vessel's operational status and providing a precise data foundation for multidimensional risk coupling analysis. Secondly, for each new energy vessel, based on its target multimodal feature vector, a large language model is used to mine the implicit risk correlation strength between multidimensional risk factors, and a risk knowledge graph is constructed based on the implicit risk correlation strength, thereby revealing the deep coupling mechanism between risk factors and establishing a semantic framework for unified multidimensional risk modeling. On this basis, based on the risk knowledge graph of each new energy vessel, a dynamic Bayesian network is used to predict the temporal evolution of risks, obtaining the single-ship risk prediction results for each new energy vessel, thus realizing the prediction of risk... The method involves quantitatively extrapolating the dynamic evolution of risks. Further, it constructs a graph structure using the individual risk prediction results and operational parameters of each new energy vessel as node features. A graph convolutional network is then used for spatial coupling modeling to obtain the cluster risk prediction results for each new energy vessel. This extends individual vessel risk to the fleet cluster level, characterizing the risk propagation and coupling effects under the mutual influence of vessels, thus forming a unified model for the spatiotemporal evolution of multidimensional risks in complex navigation scenarios. Finally, using a high-fidelity digital twin of the new energy vessel as a virtual simulation environment, and based on the risk prediction results of each cluster and a preset reward function, a reinforcement learning algorithm is used to generate the optimal vessel operation and maintenance strategy. This achieves adaptive linkage between risk extrapolation results and real-time control strategies, completing a closed-loop adaptive control from risk perception and evolution prediction to strategy generation. Therefore, this method effectively solves the closed-loop adaptive control problem from risk extrapolation to decision response in complex navigation scenarios by constructing a unified modeling mechanism for multidimensional risk coupling and linking it with real-time control strategy generation, significantly improving the safety and intelligence level of new energy vessel operation.

[0085] In one embodiment, the step of mining the strength of implicit risk associations among multidimensional risk factors using a large language model based on its target multimodal feature vector includes:

[0086] The strength of the implicit association of risk is determined by the following expression:

[0087]

[0088] in, , Let i and j be the i-th and j-th risk factor nodes in the multidimensional risk factors. The strength of the association between risk factors For large language models, For the target multimodal feature vector, As a contextualized prompt template, It is a normalized exponential function.

[0089] The technical solution expressed by this formula relies on the powerful contextual learning capabilities of a large language model to deeply explore the implicit relationships between multidimensional risk factors in new energy vessels. Specifically, by inputting a target multimodal feature vector containing the overall state of the vessel and a designed scenario-based prompt template into the large language model, the model can automatically infer the potential connections between different risk factors based on its knowledge learned from massive corpora in the field of ship operation and maintenance and its contextual understanding capabilities. These connections often cannot be discovered through traditional explicit rules or simple statistical analysis, such as the subtle influence between changes in swell and battery load in the power system. By introducing Softmax normalization, the final output correlation strength is presented as a probability value between 0 and 1, allowing the degree of coupling between different risk factors to be quantitatively expressed.

[0090] This implementation method can extract the deep coupling mechanism between the three dimensions of "environment-power-operation" risks. Guided by scenario-based prompts, the analysis process of the large language model is precisely focused on the specific application scenarios of new energy ships, avoiding the inference biases that may arise from general models. Simultaneously, relying on the contextual learning capabilities inherent in the large model, effective risk correlation mining can be achieved, significantly reducing the threshold and cost of model application. The final results of the constructed implicit risk correlation strength provide structured prior knowledge consistent with actual coupling relationships for subsequent risk knowledge graph construction and temporal evolution prediction, fundamentally improving the accuracy and comprehensiveness of modeling complex risk systems.

[0091] In one embodiment, the steps of using a dynamic Bayesian network to predict the temporal evolution of risk based on the risk knowledge graph of each new energy vessel, and obtaining the individual vessel risk prediction result for each new energy vessel, include:

[0092] The individual risk prediction result for each new energy vessel is determined according to the following expression:

[0093]

[0094] in, This represents the risk state vector of multidimensional risk factors at time t in the risk knowledge graph. The temporal feature vector representing the target multimodal feature vector. This represents the posterior probability of the risk state at time (t+1) due to the multidimensional risk factors, which is also the single-ship risk prediction result. Represents the state transition matrix. This represents a multilayer perceptron. This represents the Sigmoid function.

[0095] In this formula, risk factors from multiple dimensions, such as environment, power, and operation, can be used as nodes to construct a dynamic Bayesian network for predicting the temporal evolution of risks. During the prediction process, a state transition matrix trained based on historical risk evolution data is first used to characterize the inherent evolutionary patterns of risk factors between adjacent time points. This assumes that the current risk state is the primary determinant of the risk state at the next time point; this Markov assumption effectively simplifies the computational complexity of temporal dependencies. Simultaneously, to capture dynamic change features that cannot be reflected by the state transition matrix alone, a state correction coefficient based on the output of a multilayer perceptron is introduced. This correction coefficient takes the temporal features of the target multimodal feature vector as input and can perceive the cumulative effect of risk factors over continuous time series, such as the gradually increasing probability of battery overload under continuous surge. The state transition probability and the correction coefficient are then fused using a Sigmoid function to obtain the posterior probability of the risk state at the next time point, which closely reflects the actual operational dynamics.

[0096] This implementation, while maintaining the simplicity of the dynamic Bayesian network temporal inference framework, significantly improves the adaptability of risk prediction to actual operating conditions by introducing a temporal feature-driven correction mechanism. The state transition matrix ensures that the prediction results conform to the general laws of risk evolution, while the correction coefficients based on temporal features endow the model with the ability to perceive the cumulative effects of risk and sudden changes, enabling the prediction results to reflect complex situations such as the gradual increase in power system risk due to the continuous influence of environmental factors. The nonlinear mapping capability of the multilayer perceptron to temporal features further enhances the model's ability to capture nonlinear features in risk coupling relationships. The final single-ship risk prediction results possess both the interpretability of a probabilistic graphical model and the sensitivity of deep learning to dynamic data, providing more accurate and reliable input for subsequent cluster risk modeling and intelligent decision-making.

[0097] In one embodiment, the steps of constructing a graph structure using the individual risk prediction results and operational status parameters of each new energy vessel as node features, and employing a graph convolutional network for spatial coupling modeling to obtain the cluster risk prediction results for each new energy vessel include:

[0098] The cluster risk prediction result for each new energy vessel is determined according to the following expression:

[0099]

[0100] in, This represents the graph structure feature vector, which is also the cluster risk prediction result for each new energy vessel. Represents a graph convolutional network. This represents an adjacency matrix. The element values ​​of the adjacency matrix are assigned corresponding base values ​​according to the different intervals in which the distances between the new energy vessels fall, and the base values ​​are weighted and corrected according to the speed differences of the new energy vessels. This represents the node feature matrix, which contains the individual risk prediction results and operational status parameters for each new energy vessel. This represents the Sigmoid function. Represents the normalized adjacency matrix. This represents the activation function. and These represent the trainable weight matrices, respectively.

[0101] In this formula, the new energy vessel fleet is viewed as a dynamically changing graph structure, where each vessel is a node in the graph, and the interactions between vessels, such as navigation interference, distance correlation, and avoidance relationships, are represented as edges. The node feature matrix is ​​composed of the individual risk prediction results and operational status parameters of each new energy vessel, including the future risk evolution trend of each vessel and real-time information such as current speed, heading, and power system status. The adjacency matrix is ​​constructed based on the spatial distance and relative motion relationships between vessels; the closer the distance, the greater the mutual influence. Furthermore, by introducing speed differences to weight the elements of the adjacency matrix, vessels approaching at high speeds have stronger correlation weights, thus more realistically reflecting the intensity of risk interactions in dynamic navigation scenarios. Based on this, the graph convolutional network performs two aggregation operations. The first layer weightedly aggregates the feature information of neighboring vessels to the target node, extracting local spatial features after nonlinear transformation. The second layer further expands the receptive field, aggregating vessel features over a wider range. Finally, the graph structure feature vector of each node is output through the sigmoid function, representing the cluster risk prediction result for each new energy vessel after considering the mutual influence of other vessels in the fleet.

[0102] This implementation recalibrates the risk prediction results of individual ships, which were originally considered in isolation, within the overall framework of the fleet cluster. By explicitly modeling the spatial proximity relationships and dynamic interaction strength between ships through graph convolutional networks, the risk prediction can realistically reflect the risk propagation and superposition effects in multi-ship coexistence scenarios. The adjacency matrix, based on distance interval partitioning and speed difference weighting, ensures that the spatial coupling model has high sensitivity and accuracy in depicting the dynamic relationships between ships. Ships that are closer together and have higher relative speeds have higher weights in influencing each other's risk status, which aligns with the physical laws of high-risk scenarios such as overtaking and meeting in actual navigation. Through the hierarchical aggregation of two layers of graph convolution, the model can simultaneously capture the local influence of directly adjacent ships and the far-field effects of indirectly connected ships, fully presenting the propagation path and accumulation process of risk in the fleet network. The final cluster risk prediction result is a comprehensive output that integrates the risk status of an individual ship and the spatial coupling effects of surrounding ships, significantly improving the accuracy of risk prediction in complex waters with multiple ships coexisting. This provides a decision-making basis that conforms to actual spatial interaction relationships for subsequent development of operation and maintenance strategies involving multi-ship collaboration.

[0103] In one embodiment, the method further includes:

[0104] The risk knowledge graph is dynamically updated using the following expression:

[0105]

[0106] in, This represents the state of the risk knowledge graph at time t. Indicates the update step size. This indicates the strength of the implicit risk association among the multidimensional risk factors generated by the large language model. This indicates the confirmation result of the content generated by the large language model. This represents the cosine similarity.

[0107] This formula employs a combination of large language model extraction and manual verification to dynamically iteratively update the risk knowledge graph. During the update process, the large language model generates potential new knowledge content based on newly added navigation data, accident cases, or operation and maintenance logs. This includes new risk factor entities, implicit relationships between existing factors, and corresponding prevention and control measures. This generated content is reviewed and confirmed by senior operation and maintenance experts, forming expert feedback results corresponding to the model output. By calculating the cosine similarity between the large language model output and the expert feedback results, the consistency between the model-generated content and actual domain knowledge is quantitatively evaluated. Based on this, the risk knowledge graph is incrementally corrected according to the preset update step size and specific knowledge update actions, enabling the knowledge graph to continuously evolve with the emergence of new scenarios and cases.

[0108] This implementation effectively integrates the automated extraction capabilities of large language models with the domain experience judgment of human experts. It leverages the efficiency advantage of large models in mining implicit relationships from massive amounts of text while ensuring the accuracy and reliability of newly added knowledge through manual verification, avoiding potential errors introduced by fully automated updates. The introduction of cosine similarity provides a reasonable constraint on knowledge updates; updates only have a significant impact on the knowledge graph when the model output and expert confirmation are semantically highly similar, thus ensuring the stability and gradual nature of knowledge evolution. Through a dynamic update mechanism, the risk knowledge graph can continuously absorb the latest navigation practices and failure cases, constantly enriching and refining the network of relationships between risk factors, ensuring it always aligns with the actual changes in the field of new energy vessel operation and maintenance. The resulting dynamically evolving knowledge graph provides a timely knowledge foundation for subsequent risk prediction and intelligent decision-making, effectively improving the model's adaptability to new scenarios and the timeliness of risk identification.

[0109] In one embodiment, the method further includes:

[0110] By using federated learning and collaborating local data from multiple new energy ship data sources, the parameters of dynamic Bayesian networks and graph convolutional networks are trained, and the general model parameters of dynamic Bayesian networks and graph convolutional networks are obtained.

[0111] Scene data of the target navigation scenario is acquired, and the general model parameters of the dynamic Bayesian network and the graph convolutional network are fine-tuned through transfer learning to obtain the adapted model parameters of the dynamic Bayesian network and the graph convolutional network in the target navigation scenario.

[0112] Specifically, by organically combining federated learning and transfer learning, the two core challenges of data silos and scenario differences in risk modeling for new energy vessels are systematically addressed. In the federated learning phase, homomorphic encryption technology is used to construct a secure collaborative training framework, enabling multiple data holders, such as shipping companies or ports, to jointly participate in the training of the global risk model without leaving their local data. Each participant calculates a local loss function based on its own data, with the objective function being to minimize the sum of the weighted local losses of all participants, i.e.:

[0113] ,in

[0114] Where K represents the number of entities participating in federal training. Let N be the amount of data for the k-th party, and N be the total amount of data for all participating parties. Let θ be the k-th local loss function. The local gradients of each subject are aggregated using a federated averaging algorithm, and the global model parameters θ, i.e., the general model parameters of the dynamic Bayesian network and the graph convolutional network, are iteratively optimized to generate a general risk model. This process fully utilizes the rich navigation scenarios and risk patterns contained in multi-source data, enabling the trained general model parameters to have stronger generalization ability and effectively avoiding model bias that may be caused by the limitations of single-subject data.

[0115] After obtaining the general model parameters, domain adaptation techniques are used for transfer learning optimization to address the data distribution differences between the target navigation scenario (e.g., inland waterways or narrow channels) and the source scenario (e.g., open coastal areas). An adaptation loss function is introduced to quantify the difference in feature distribution between the source and target scenarios, and gradients are calculated using a small amount of labeled data from the target scenario to fine-tune the general model parameters. The fine-tuning formula is as follows:

[0116]

[0117] in, For the target scene model parameters, Pre-trained parameters for the source scene The learning rate is set at 0.001 to 0.01, balancing source knowledge retention with target adaptation (preferably 0.005). To adapt to the loss function, Provide a small amount of labeled data for the target scene (no more than 10%). During fine-tuning, set a reasonable learning rate to shift the model parameters towards the characteristics of the target scene while retaining the pre-trained knowledge of the source scene, ultimately generating model parameters adapted to the target scene. .

[0118] In this implementation, the federated learning mechanism effectively breaks down data barriers between different shipping companies and ports, achieving cross-entity data collaboration value while ensuring commercial privacy and data security. This enables the model to learn broader risk evolution patterns and spatial coupling modes, significantly improving its generalization ability across different fleets and waterways. The transfer learning mechanism addresses the performance degradation issue when general models are directly applied to specific scenarios. It allows for rapid model adaptation using target scenario data, significantly reducing data annotation costs and model deployment cycles in new scenarios. This combined approach, moving from global generalization to local adaptation, ensures the model's basic accuracy while providing fine-tuning capabilities for specific scenarios. Ultimately, this allows dynamic Bayesian networks and graph convolutional networks to maintain stable predictive performance across diverse navigation scenarios, including nearshore, inland waterway, and coastal waterways.

[0119] In one embodiment, the method further includes:

[0120] The SHAP algorithm was used to determine the importance values ​​of various risk factors in the multidimensional risk factors.

[0121] The importance values ​​of various risk factors and the risk knowledge graph are input into a domain-fine-tuned large language model to generate natural language explanations and operational suggestions that fit the operation and maintenance scenario.

[0122] This embodiment utilizes interpretable artificial intelligence technology to transform the black-box reasoning process of intelligent models into logic and suggestions that humans can understand. In the risk factor importance quantification stage, the SHAP algorithm, based on game theory principles, calculates the marginal contribution of each risk factor to the prediction result. Its mathematical expression is as follows:

[0123]

[0124] This formula calculates the model prediction function before and after adding the i-th risk factor by iterating through all feature subsets S that do not contain the i-th risk factor. The change in the factor is calculated, and a weighted average is applied to different subsets to obtain the fair contribution of the factor to the prediction results. The value ranges from -1 to 1, with positive values ​​indicating that the factor promotes risk and negative values ​​indicating that it inhibits risk. Through this quantification process, we can calculate the local contribution of each factor for a single risk event, or aggregate a large number of samples from a global perspective to identify the core risk sources in different navigation scenarios.

[0125] In the natural language interpretation and generation stage, the importance values ​​calculated by the SHAP algorithm are combined with the relational logic stored in the risk knowledge graph to form a comprehensive input containing quantitative evidence and logical links. The generated model takes the following form:

[0126]

[0127] In this formula, ⊕ represents the feature concatenation operation, which combines the SHAP quantization result vector F_SHAP with the risk knowledge graph feature vector. Fusion represents a joint representation; Template is a scenario-based prompt template designed according to elements such as scenario, risk level, core cause, quantitative contribution, and prevention and control suggestions; G is a large language model fine-tuned in the field of ship operation and maintenance, whose role is to transform technical feature inputs into natural language descriptions that fit operation and maintenance practices, and generate executable explanatory text containing specific operation suggestions.

[0128] The SHAP algorithm is employed to determine the importance values ​​of various risk factors within a multidimensional risk framework. These importance values, along with a risk knowledge graph, are then input into a domain-fine-tuned large language model to generate natural language explanations and operational suggestions. The aim is to transform the reasoning process of a black-box model into quantifiable attribution results, and further into human-understandable natural language expressions. Based on game theory principles, the SHAP algorithm fairly allocates the marginal contribution of each factor to the prediction result, enabling precise quantification of the promoting or inhibiting effect of each risk factor on the risk state, thus solving the problem of untraceable internal logic within the intelligent model. The quantified importance values ​​are integrated with the relationships stored in the risk knowledge graph, providing structured knowledge support for subsequent explanation generation. Building upon this, the domain-fine-tuned large language model transforms technical numerical features and logical links into natural language outputs tailored to the operational scenario, with the generated explanations and suggestions directly linked to specific operational actions. The SHAP algorithm was used to fairly quantify the contribution of risk factors, making the risk reasoning process interpretable and traceable. The quantification results and knowledge graph logic were transformed into easy-to-understand language through a large language model, which effectively broke down the cognitive gap between operation and maintenance personnel and intelligent models, significantly improved the credibility and adoptability of decision-making suggestions, and finally solved the problem that traditional intelligent models are difficult to implement in actual operation and maintenance due to the lack of transparency in the reasoning process.

[0129] To facilitate understanding of the scheme in this application, specific examples are provided below.

[0130] This embodiment uses an inland waterway battery-powered cargo ship as the application object, employs LLaMA-7B as the basic large model, and constructs the core model using GCN, DBN, and PPO algorithms. The hardware environment consists of a GPU (NVIDIA A100, 80GB VRAM) and a CPU (Intel Xeon Platinum 8470C, 24 cores and 48 threads), while the software environment uses PyTorch 2.0 and TensorFlow 2.10. The specific steps, supporting mathematical models, and parameter descriptions are as follows:

[0131] Step 1: Construct a multi-source knowledge foundation.

[0132] This step relies on the S11 multimodal data fusion formula and the S12 risk association and knowledge graph update model to achieve a structured combination of data and knowledge, with parameter values ​​adapted to the characteristics of inland waterway pure battery cargo ships.

[0133] S11: Multimodal Data Fusion Practice and Formula Parameter Explanation: Collect sensor data from the cargo ship's power system (5000+ sets of time-series data, sampling frequency 1Hz, including 12 types of parameters such as battery temperature, voltage, and motor speed), AIS trajectory data (inland waterway navigation trajectory for the past 3 months, time accuracy to the second, including 6 types of parameters such as latitude and longitude, speed, and heading), meteorological and hydrological data (inland waterway water level, current speed, and visibility, updated every 10 minutes), 800+ maintenance logs, 200+ inland waterway vessel accident cases, and the CCS "Inspection Guidelines for Pure Battery-Powered Inland Waterway Vessels" standard text. Use the following multimodal data fusion formula:

[0134]

[0135] In the formula, , , , Based on the characteristics of inland waterway scenarios—the stability of the propulsion system has the greatest impact on navigation safety, therefore... Take the maximum value; inland waterways are narrow, and AIS spatial data is crucial for obstacle avoidance risk modeling. Next.

[0136] Feature vector: 768-dimensional time-series features were generated using Time2Vec encoding, with a time window set to 60 seconds, to capture short-term trends in battery voltage fluctuations and motor temperature changes. 768-dimensional spatial features are generated by GCN encoding, with input spatial parameters such as latitude and longitude, channel width, and distance from shoals; 768-dimensional text features were extracted using BERT-base, focusing on fault descriptions and accident case triggering keywords in maintenance logs; 768-dimensional environmental features are generated through VLM encoding, with core associations including inland river water level (safe threshold 1.2-3.0 meters) and water flow velocity (safe threshold ≤2 m / s).

[0137] In the data preprocessing stage, the 3σ criterion was used to remove sensor outliers (removal rate of approximately 3.2%, threshold range of mean ± 3 times standard deviation, with separate calibration for key parameters such as battery voltage and motor temperature). The KNN algorithm (neighbor number k=5, distance metric using Euclidean distance) was used to complete missing values. Semantic matching (based on the TF-IDF algorithm, similarity threshold ≥ 0.7) and temporal calibration (timestamp error ≤ 1 second) were used to achieve accurate alignment of multi-source data, ultimately forming a standardized dataset with dimensions of 1024×8000. (1024-dimensional features, 8000 samples).

[0138] S12: Risk Association Mining and Knowledge Graph Construction Parameter Explanation: Implicit associations are mined using the LLaMA-7B large model (LoRA fine-tuning parameters: rank r=8, learning rate 5e-5, 500 iterations), with the core relying on the risk feature association model:

[0139]

[0140] In the formula, , There are 25 risk factor nodes (8 environmental nodes: shoals, water flow speed, etc.; 10 power nodes: battery SOC, motor speed, etc.; 7 operational nodes: crew operation, route planning, etc.). The Prompt template is set to "Based on the scenario of a 1000-ton inland waterway pure battery cargo ship, analyze the correlation between {environmental factors} / {power factors} / {operational factors} and {failure type}, output the correlation strength in the 0-1 interval, and retain two decimal places".

[0141] Output results: Implicit associations were found, such as "shallow water + battery squeezing → short circuit risk" with a correlation strength of 0.81 and "turbulent water flow → speed loss → collision risk" with a correlation strength of 0.75, forming a 25×25-dimensional correlation strength matrix.

[0142] The knowledge graph is constructed using a method of "large model extraction + verification by two senior operations and maintenance experts," and iteration is achieved through dynamic model updates. Parameter settings (To balance the speed and stability of knowledge updates and avoid new data from disturbing the original core relationships), Sim(·) uses cosine similarity calculation (if the threshold is ≥0.85, the updated content is retained) to finally construct a risk knowledge graph containing more than 1,200 entities and more than 3,500 relationships. The entity attributes include key information such as parameter thresholds and prevention and control priorities.

[0143] Step 2: Construct a dynamic coupling model.

[0144] By integrating the S21 DBN temporal model and the S22 GCN spatial model, the temporal-spatial dual-dimensional coupled quantification of inland waterway fleet risk is achieved, with parameters adapted to the scenario of three 1,000-ton cargo ships sailing in formation.

[0145] S21: DBN Temporal Dynamic Risk Modeling Practice and Parameter Explanation: Construct a DBN model, set 25 risk nodes (corresponding one-to-one with knowledge graph nodes), time series step size T=90 (corresponding to a 90-second time series window, adapting to the response cycle of sudden risks in inland waterways), based on the state transition formula:

[0146]

[0147] State transition matrix : Dimension is 25×25, trained based on the risk evolution data of inland river pure battery cargo ships in the past year. For example, the transition probability of "water level below 1.2 meters → risk of propeller stranding" is 0.45, and the transition probability of "battery SOC below 30% → risk of power interruption" is 0.62.

[0148] Temporal features : Use Transformer to extract the temporal data features within a 90-second window, with an output dimension of 512; MLP is set to 2 layers (input 512 dimensions, hidden layer 256 dimensions, output 25 dimensions), σ is the Sigmoid function, and the correction coefficient is normalized to [0,1] to capture cumulative effects such as the continuous action of surges and the attenuation of battery performance.

[0149] The model is trained using the Adam optimizer (learning rate 1e-3, 1000 iterations), and the training loss converges within 0.03, enabling accurate prediction of "when the inland river water level is below the safety threshold (1.2 meters) → the risk of propeller stranding within 30 minutes increases from 12% to 52%".

[0150] S22: Practical operation and parameter description of GNN spatial risk coupling modeling: Consider 3 ships in the fleet where the cargo ship is located as a graph structure (nodes are individual ships, edges are navigation interference and distance association), and use GCN to construct a spatial coupling model:

[0151]

[0152] Adjacency matrix A: Dimension is 3×3, quantified based on navigation distance (threshold 50 meters) and interference intensity. When the distance d ≤ 20 meters, the element value is 0.8; when 20 meters < d ≤ 50 meters, it is 0.3; when d > 50 meters, it is 0; the interference intensity is calculated by combining the speed difference and the course angle, and when the speed difference ≥ 3 knots, a weight of 0.2 is superimposed.

[0153] Normalized adjacency matrix : Through Calculation ( is the adjacency matrix with self-loops added, is 's degree matrix); (input 25 dimensions, output 128 dimensions), (input 128 dimensions, output 25 dimensions) are trainable weight matrices, and are trained through the fleet navigation data to adapt to the spatial interaction characteristics of the inland river channel.

[0154] The model outputs the coupling coefficient of "single-ship risk → cluster risk". For example, when the leading ship slows down to avoid a shoal, the coupling coefficient of the following ship's risk of rear-ending is 0.78, quantifying the superimposed impact of spatial interaction on risk.

[0155] Step 3: Cross-scenario generalization adaptation.

[0156] Based on the S31 federated learning and S32 transfer learning models, the model for inland waterway scenarios is adapted to coastal narrow waterway scenarios, taking into account both data privacy and generalization accuracy.

[0157] S31: Federated Learning Cross-Agent Collaborative Modeling Practice and Parameter Explanation: Using data from three inland waterway shipping companies, a collaborative training framework is constructed using homomorphic encryption technology (Paillier encryption algorithm, 1024-bit key length). The objective function is:

[0158] ,in

[0159] In the formula, K=3 (3 shipping companies). , , (Sample size for each company), N=8000; θ is the global model parameter (dimension is consistent with DBN and GCN model parameters). These are the parameters of the k-th local model.

[0160] Local loss function The model employs cross-entropy loss to quantify the deviation between the predicted risk level and the actual risk status. ( This is an actual risk label. (For predicting probabilities).

[0161] The gradient is aggregated using a federated averaging algorithm (500 iterations, 10-round interval). The local loss functions of all participants converge to within 0.02. Homomorphic encryption is used to ensure that the original data is not leaked, and a general risk model is generated.

[0162] S32: Practical Exercises and Parameter Explanations for Transfer Learning Scene Adaptation and Optimization: A pre-trained model for an inland river scene is transferred to a coastal narrow waterway scene, and fine-tuned using domain adaptation techniques. The model is as follows:

[0163]

[0164] In the formula, the learning rate η = 0.005 (to balance the retention of source knowledge with the target adaptation, and avoid excessive fine-tuning that could lead to the loss of knowledge in the inland river scenario). The maximum mean difference (MMD) loss is used to quantify the difference in feature distribution between the source scene (inland river) and the target scene (coastal narrow waterway), and the kernel function is a Gaussian kernel (bandwidth σ=0.1).

[0165] Target scene data We selected 8% of the labeled data from coastal narrow waterways (640 samples in total), including scenario-specific parameters such as narrow waterway navigation density and tidal current speed. After fine-tuning, the model's risk prediction accuracy reached 92%, meeting the needs of coastal narrow waterway scenarios.

[0166] Step 4: Generate interpretable reasoning.

[0167] Based on the S41 SHAP algorithm and the S42 natural language generation model, we can realize the quantitative interpretation of risk reasoning results and output practical suggestions.

[0168] S41: Practical Operation and Parameter Explanation for Quantifying the Importance of Risk Factors: The SHAP algorithm is used to quantify the contribution of risk factors. The model is as follows:

[0169]

[0170] In the formula, K = 25 (total number of risk factors), and S is a subset of risk factors. This is the prediction function for the DBN-GNN fusion model (input feature vector, output risk level probability). The value range is [-1, 1]. Positive values ​​promote risk, while negative values ​​suppress risk. The calculation is performed using the SHAP Python library (version 0.41.0) and the number of samples is 1000.

[0171] Output results: In the inland waterway scenario, the top three risk factors are shoal distribution (SHAP value 0.82), battery SOC (SHAP value 0.65), and crew operation (SHAP value 0.58). It is clear that the core cause of the grounding risk is shoal distribution, which accounts for more than 40% of the risk.

[0172] S42: Practical Operation and Parameter Explanation of Natural Language Interpretation Generation: Based on the domain-fine-tuned LLaMA-7B model, and integrating SHAP results with knowledge graph logic, the generated model is as follows:

[0173]

[0174] Feature concatenation: Concatenate the SHAP value vector (25-dimensional) with the graph structure feature vector (25-dimensional) to form a 50-dimensional feature, and input it into the large model; The template is fixed as "Current {scenario} risk level is {level}, the core cause is {factor 1} (SHAP value {value 1}) and {factor 2} (SHAP value {value 2}), and the suggested {measure 1} and {measure 2} are based on {knowledge graph association}".

[0175] In the formula, the batch size of the large model fine-tuning is 8, and the iteration is 300 rounds. The generated natural language suggestions are close to the actual operation of inland waterway maintenance. For example, "At present, there are many shallow shoals in the inland waterway. The battery SOC is 82%. It is recommended to reduce the speed to 8 knots and sail along the deep water route (offset ≥ 50 meters) to avoid the risk of propeller entanglement and grounding. At the same time, check the battery charging and discharging status every 15 minutes and prevent and control the risk based on the transmission link of 'shallow shoals → battery compression → short circuit risk'."

[0176] Step 5: Digital Twins and Decision Optimization.

[0177] A high-fidelity digital twin is constructed and combined with the S52 PPO reinforcement learning algorithm to form a "prediction-decision-feedback" closed loop, with parameters adapted to the physical characteristics and operation and maintenance requirements of a 1,000-ton inland waterway pure battery cargo ship.

[0178] S51: Practical Guide and Parameter Explanation for Building a High-Fidelity Digital Twin: Combining four sub-models—geometric, physical, behavioral, and environmental—with accuracy constraints met:

[0179]

[0180] In the formula, The digital twin's state parameters include 100 key parameters, covering hull attitude, battery voltage, motor speed, channel water level, etc. The parameters are measured in real time by physical ships (collected in real time by sensors at a sampling frequency of 1Hz); error calculation uses Euclidean distance, and the time window is set to 10 seconds to ensure short-term dynamic consistency.

[0181] Model calibration: The physical model was calibrated using 100 sets of measured data. The fluid dynamics model was simulated using STAR-CCM+, and the battery model was built based on the equivalent circuit model. The final simulation error was controlled within 4.2%, which can map the ship's operating status and changes in the inland waterway environment in real time.

[0182] S52: Practical Exercises and Parameter Explanation for Reinforcement Learning Decision Optimization: Using a digital twin as the virtual environment, the agent is trained using the PPO algorithm, with the reward function as follows:

[0183]

[0184] Weights and sub-reward functions: , , Safety should be the top priority. (Risk level 0-1, corresponding reward 1-0) (Threshold 0.8-1.2, rewards are reduced if outside the range). .

[0185] PPO algorithm parameters: pruning coefficient ε=0.2, discount factor γ=0.99, learning rate 3e-4, model converges after 1000 iterations, action space includes speed adjustment (5-12 knots), course deviation (0-100 meters), and inspection frequency adjustment (5-30 minutes), state space includes 25 risk node states and 100 twin parameters.

[0186] The training generated an optimal strategy of "shoal warning → adjust course (offset ≥ 50 meters) + reduce speed to 8 knots". The decision response delay was 850ms. After one month of real ship testing, the occurrence rate of the ship's core risks (grounding, battery failure, collision) was reduced by 20% and the operation and maintenance efficiency was improved by 15%, which verified the practicality and reliability of the model.

[0187] The intelligent decision-making device based on multi-dimensional risk coupling modeling of new energy ships, provided in the embodiments of this application, is described below. The intelligent decision-making device based on multi-dimensional risk coupling modeling of new energy ships described below can be referred to in correspondence with the intelligent decision-making method based on multi-dimensional risk coupling modeling of new energy ships described above. Figure 2 As shown, this application provides an intelligent decision-making device based on multi-dimensional risk coupling modeling of new energy ships. The device includes:

[0188] The multimodal data fusion module 201 is used to fuse the multimodal data of each new energy vessel in the new energy vessel fleet to obtain the target multimodal feature vector of each new energy vessel.

[0189] The risk knowledge graph construction module 202 is used to mine the implicit risk correlation strength between multi-dimensional risk factors based on the target multimodal feature vector of each new energy ship, and construct a risk knowledge graph according to the implicit risk correlation strength.

[0190] The single-ship risk prediction result determination module 203 is used to predict the risk time series evolution based on the risk knowledge graph of each new energy vessel and to obtain the single-ship risk prediction result for each new energy vessel.

[0191] The cluster risk prediction result determination module 204 is used to construct a graph structure using the individual ship risk prediction results and operating status parameters of each new energy ship as node features, and to perform spatial coupling modeling using a graph convolutional network to obtain the cluster risk prediction result of each new energy ship.

[0192] The optimal ship operation and maintenance strategy generation module 205 is used to generate the optimal ship operation and maintenance strategy based on the risk prediction results of each cluster and the preset reward function, using a high-fidelity digital twin of a new energy ship as a virtual simulation environment and a reinforcement learning algorithm.

[0193] In one embodiment, the risk knowledge graph construction module 202 includes:

[0194] The risk implicit association strength determination unit is used to determine the risk implicit association strength according to the following expression:

[0195]

[0196] in, , Let i and j be the i-th and j-th risk factor nodes in the multidimensional risk factors. The strength of the association between risk factors For large language models, For the target multimodal feature vector, As a contextualized prompt template, It is a normalized exponential function.

[0197] In one embodiment, the single-ship risk prediction result determination module 203 includes:

[0198] The single-ship risk prediction result determination unit is used to determine the single-ship risk prediction result for each new energy vessel according to the following expression:

[0199]

[0200] in, This represents the risk state vector of multidimensional risk factors at time t in the risk knowledge graph. The temporal feature vector representing the target multimodal feature vector. This represents the posterior probability of the risk state at time (t+1) due to the multidimensional risk factors, which is also the single-ship risk prediction result. Represents the state transition matrix. This represents a multilayer perceptron. This represents the Sigmoid function.

[0201] In one embodiment, the cluster risk prediction result determination module 204 includes:

[0202] The cluster risk prediction result determination unit is used to determine the cluster risk prediction result for each new energy vessel according to the following expression:

[0203]

[0204] in, This represents the graph structure feature vector, which is also the cluster risk prediction result for each new energy vessel. Represents a graph convolutional network. This represents an adjacency matrix. The element values ​​of the adjacency matrix are assigned corresponding base values ​​according to the different intervals in which the distances between the new energy vessels fall, and the base values ​​are weighted and corrected according to the speed differences of the new energy vessels. This represents the node feature matrix, which contains the individual risk prediction results and operational status parameters for each new energy vessel. This represents the Sigmoid function. Represents the normalized adjacency matrix. This represents the activation function. and These represent the trainable weight matrices, respectively.

[0205] In one embodiment, the apparatus further includes:

[0206] The risk knowledge graph dynamic update module is used to dynamically update the risk knowledge graph using the following expression:

[0207]

[0208] in, This represents the state of the risk knowledge graph at time t. Indicates the update step size. This indicates the strength of the implicit risk association among the multidimensional risk factors generated by the large language model. This indicates the confirmation result of the content generated by the large language model. This represents the cosine similarity.

[0209] In one embodiment, the apparatus further includes:

[0210] The federated learning module is used to train the parameters of dynamic Bayesian networks and graph convolutional networks by leveraging local data from multiple new energy ship data sources through federated learning, thereby obtaining the general model parameters of dynamic Bayesian networks and graph convolutional networks.

[0211] The transfer learning module is used to acquire scene data of the target navigation scenario and fine-tune the general model parameters of the dynamic Bayesian network and the graph convolutional network through transfer learning to obtain the adapted model parameters of the dynamic Bayesian network and the graph convolutional network in the target navigation scenario.

[0212] In one embodiment, the apparatus further includes:

[0213] The importance value determination module is used to determine the importance values ​​of various risk factors in multidimensional risk factors using the SHAP algorithm;

[0214] The natural language interpretation and generation module is used to input the importance values ​​of various risk factors and the risk knowledge graph into the domain-fine-tuned large language model to generate natural language interpretations and operation suggestions that fit the operation and maintenance scenario.

[0215] In one embodiment, this application also provides a storage medium storing computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the steps of the intelligent decision-making method based on multidimensional risk coupling modeling of new energy ships as described in any of the above embodiments.

[0216] In one embodiment, this application also provides a computer device storing computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the steps of the intelligent decision-making method based on multidimensional risk coupling modeling of new energy ships as described in any of the above embodiments.

[0217] Indicatively, such as Figure 3 As shown, Figure 3 This is a schematic diagram of the internal structure of a computer device 300 provided in an embodiment of this application. The computer device 300 can be provided as a server. (Refer to...) Figure 3 The computer device 300 includes a processing component 302, which further includes one or more processors, and memory resources represented by memory 301 for storing instructions, such as application programs, that can be executed by the processing component 302. The application programs stored in memory 301 may include one or more modules, each corresponding to a set of instructions. Furthermore, the processing component 302 is configured to execute instructions to perform the intelligent decision-making method based on multi-dimensional risk coupling modeling of new energy ships according to any of the above embodiments.

[0218] The computer device 300 may also include a power supply component 303 configured to perform power management of the computer device 300, a wired or wireless network interface 304 configured to connect the computer device 300 to a network, and an input / output (I / O) interface 305. The computer device 300 may operate on an operating system stored in memory 301, such as Windows Server™, Mac OS X™, Unix™, Linux™, Free BSD™, or similar.

[0219] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0220] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. In this document, "a," "an," "the," "the," and "its" may also include plural forms unless the context clearly indicates otherwise. "Multiple" refers to at least two, such as 2, 3, 5, or 8, etc. "And / or" includes any and all combinations of the related listed items.

[0221] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

[0222] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An intelligent decision-making method based on multi-dimensional risk coupling modeling of new energy ships, characterized in that, The method includes: In the new energy vessel fleet, the multimodal data of each new energy vessel are fused to obtain the target multimodal feature vector of each new energy vessel; For each of the new energy vessels, based on its target multimodal feature vector, a large language model is used to mine the strength of the implicit risk association between multidimensional risk factors, and a risk knowledge graph is constructed according to the strength of the implicit risk association. Based on the risk knowledge graph of each of the new energy vessels, a dynamic Bayesian network is used to predict the temporal evolution of risk, and the single-ship risk prediction results of each of the new energy vessels are obtained. A graph structure is constructed using the individual risk prediction results and operating status parameters of each new energy vessel as node features. Spatial coupling modeling is performed using a graph convolutional network to obtain the cluster risk prediction results of each new energy vessel. Using a high-fidelity digital twin of a new energy vessel as a virtual simulation environment, and based on the risk prediction results of each cluster and the preset reward function, a reinforcement learning algorithm is used to generate the optimal vessel operation and maintenance strategy.

2. The intelligent decision-making method based on multi-dimensional risk coupling modeling of new energy ships according to claim 1, characterized in that, The step of mining the strength of implicit risk associations among multidimensional risk factors using a large language model based on its target multimodal feature vector includes: The strength of the implicit association of the risk is determined according to the following expression: in, , For the i-th and j-th risk factor nodes in the multidimensional risk factors, The strength of the association between risk factors For large language models, The target multimodal feature vector, As a contextualized prompt template, It is a normalized exponential function.

3. The intelligent decision-making method based on multi-dimensional risk coupling modeling of new energy ships according to claim 1, characterized in that, The step of using a dynamic Bayesian network to predict the temporal evolution of risk based on the risk knowledge graph of each new energy vessel, and obtaining the single-ship risk prediction result for each new energy vessel, includes: The individual risk prediction result for each of the aforementioned new energy vessels is determined according to the following expression: in, This represents the risk state vector of the multidimensional risk factor at time t in the risk knowledge graph. This represents the temporal feature vector of the target multimodal feature vector. This represents the posterior probability of the risk state of the multidimensional risk factors at time (t+1), which is also the single-ship risk prediction result. Represents the state transition matrix. This represents a multilayer perceptron. This represents the Sigmoid function.

4. The intelligent decision-making method based on multi-dimensional risk coupling modeling of new energy ships according to claim 1, characterized in that, The step of constructing a graph structure using the individual risk prediction results and operating status parameters of each new energy vessel as node features, and using a graph convolutional network for spatial coupling modeling to obtain the cluster risk prediction results of each new energy vessel includes: The cluster risk prediction result for each of the aforementioned new energy vessels is determined according to the following expression: in, The graph structure feature vector represents the cluster risk prediction result for each of the aforementioned new energy vessels. This represents a graph convolutional network. This represents an adjacency matrix, where the element values ​​are assigned corresponding base values ​​based on the different intervals in which the distances between the new energy vessels fall, and the base values ​​are weighted and corrected based on the speed differences between the new energy vessels. This represents the node feature matrix, which is also the individual ship risk prediction results and operating status parameters of each of the aforementioned new energy vessels. This represents the Sigmoid function. Represents the normalized adjacency matrix. This represents the activation function. and These represent the trainable weight matrices, respectively.

5. The intelligent decision-making method based on multi-dimensional risk coupling modeling of new energy ships according to claim 1, characterized in that, The method further includes: The risk knowledge graph is dynamically updated using the following expression: in, This represents the state of the risk knowledge graph at time t. Indicates the update step size. This indicates the strength of the implicit risk association among the multidimensional risk factors generated by the large language model. This indicates the confirmation result of the content generated by the large language model. This represents the cosine similarity.

6. The intelligent decision-making method based on multi-dimensional risk coupling modeling of new energy ships according to claim 1, characterized in that, The method further includes: By using federated learning and collaborating with local data from multiple new energy ship data sources, the parameters of the dynamic Bayesian network and the graph convolutional network are trained to obtain the general model parameters of the dynamic Bayesian network and the graph convolutional network. Scene data of the target navigation scenario is acquired, and the general model parameters of the dynamic Bayesian network and the graph convolutional network are fine-tuned through transfer learning to obtain the adapted model parameters of the dynamic Bayesian network and the graph convolutional network in the target navigation scenario.

7. The intelligent decision-making method based on multi-dimensional risk coupling modeling of new energy ships according to claim 1, characterized in that, The method further includes: The importance values ​​of each type of risk factor in the multidimensional risk factors are determined using the SHAP algorithm. The importance values ​​of each of the aforementioned risk factors and the risk knowledge graph are input into a domain-fine-tuned large language model to generate natural language explanations and operational suggestions that fit the operation and maintenance scenario.

8. An intelligent decision-making device based on multi-dimensional risk coupling modeling of new energy ships, characterized in that, The device includes: The multimodal data fusion module is used to fuse the multimodal data of each new energy vessel in the new energy vessel fleet to obtain the target multimodal feature vector of each new energy vessel. The risk knowledge graph construction module is used to, for each of the new energy vessels, mine the implicit risk correlation strength between multidimensional risk factors based on its target multimodal feature vector using a large language model, and construct a risk knowledge graph based on the implicit risk correlation strength. The single-ship risk prediction result determination module is used to perform risk time series evolution prediction based on the risk knowledge graph of each new energy vessel and a dynamic Bayesian network to obtain the single-ship risk prediction result for each new energy vessel. The cluster risk prediction result determination module is used to construct a graph structure using the individual ship risk prediction results and operating status parameters of each new energy vessel as node features, and to perform spatial coupling modeling using a graph convolutional network to obtain the cluster risk prediction result of each new energy vessel. The optimal ship operation and maintenance strategy generation module is used to generate the optimal ship operation and maintenance strategy based on the risk prediction results of each cluster and the preset reward function, using a high-fidelity digital twin of a new energy ship as a virtual simulation environment and a reinforcement learning algorithm.

9. A storage medium, characterized in that: The storage medium stores computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the steps of the intelligent decision-making method based on multidimensional risk coupling modeling of new energy ships as described in any one of claims 1 to 7.

10. A computer device, characterized in that, include: One or more processors, and memory; The memory stores computer-readable instructions, which, when executed by the one or more processors, perform the steps of the intelligent decision-making method based on multidimensional risk coupling modeling of new energy ships as described in any one of claims 1 to 7.

Images